In a standard search engine, every query is an island. If you search for "The Eiffel Tower" and then follow up with "How tall is it?", the engine needs to understand that "it" refers to the tower. In conversational information seeking (CIS), this problem is magnified across dozens of turns. Users naturally use pronouns, omit context, and expect the system to remember the "thread" of the conversation.
The central challenge in this field is context drift: as a conversation progresses, the search system often loses track of the original intent, leading to irrelevant results. Recent research has converged on a powerful, intuitive solution: the first turn of a conversation is not just the beginning; it is a persistent semantic anchor that defines the entire session.
The Power of the First Turn
When users engage in multi-turn search, they typically explore a single core topic in depth before shifting focus. Because of this behavior, the very first utterance often contains the most explicit description of the user's information need.
Mohammad Aliannejadi, Esteban A. Ríssola, and their colleagues demonstrated in Predicting Relevant Conversation Turns for Improved Retrieval in Multi-Turn Conversational Search that the initial turn serves as a primary context anchor in more than half of all cases. They found that simply including the first turn as a constant contextual prefix for every subsequent query consistently improves retrieval performance.
This "first-turn anchoring" heuristic works because it preserves thematic continuity. Even as a user asks specific, narrow follow-up questions, the system remains "tethered" to the overarching premise established at the start. This is further supported by the work of Ming Jiang and Junli Wang in A Speaker-aware Parallel Hierarchical Attentive Encoder-Decoder Model for Multi-turn Dialogue Generation, where attention visualizations revealed that successful models frequently refer back to the initial input to maintain global topic consistency.
Query Reformulation and Selective Enrichment
If the first turn is the anchor, how do we actually use it to improve search? The dominant approach is query reformulation: rewriting a vague follow-up (e.g., "What about its history?") into a standalone, searchable query (e.g., "What is the history of the Eiffel Tower?").
In Harnessing Evolution of Multi-Turn Conversations for Effective Answer Retrieval, Mohammad Aliannejadi and his team introduced BERT-CUR, a model that identifies which specific historical turns are essential for the current question. By automatically reformulating queries with this historical context, they achieved a 38% relative improvement in answer retrieval.
However, simply appending all previous turns creates "conversational noise." To solve this, Ida Mele and Nicola Tonellotto proposed a strategy of topical enrichment in Topical Enrichment of Conversational Search Utterances. Instead of using raw text, they use Named Entity Recognition to extract high-signal concepts from previous turns and system responses. This ensures that only high-salience entities—the "who" and the "what"—are propagated forward, filtering out irrelevant conversational filler.
From Keywords to Semantic Representations
While early conversational search relied on keyword matching (like BM25), the field is shifting toward Dense Passage Retrieval (DPR). In this paradigm, queries and documents are converted into high-dimensional vectors where proximity represents semantic similarity.
Pierre McWhannel and Nicole Yan explored this in Dense Passage Retrieval in Conversational Search. They combined a GPT-2-based reformulator with a dual-encoder dense retriever. Their findings reinforce the anchoring intuition: incorporating the initial utterance as a persistent prefix to subsequent queries provides a substantial boost in retrieval performance. By maintaining the core topic context in the vector space, the system can "understand" the relationship between the first turn and a vague follow-up even if they share no common keywords.
Summary of the Intuition
The evolution of this research direction suggests a clear mental model for conversational search:
- The First Turn is King: It establishes the semantic boundaries of the session.
- Anchoring: By treating the first turn as a persistent context, we prevent the "drift" that occurs as conversations become more specific.
- Signal over Noise: Effective systems don't just remember everything; they selectively enrich current queries with named entities and semantic anchors from the past.